fix(llm): raise OpenRouter Qwen3 verification budget

[r266-tech]

Summary

• keep the normal OpenAI-compatible startup verification budget at 100 tokens
• raise it to 512 only for OpenRouter qwen/qwen3... model IDs, where hidden reasoning can consume the small budget before visible ok content is emitted
• add boundary coverage for Qwen3 positives plus Qwen2.5, near-miss aliases, and ordinary custom gateways staying on the default budget

Fixes #2630.

Tests

• python3 -m py_compile hindsight-api-slim/hindsight_api/engine/providers/openai_compatible_llm.py hindsight-api-slim/tests/test_openai_compatible_response_hardening.py

Focused pytest could not run in this local sparse checkout because tests/conftest.py imports project dependencies not installed in the cron environment (asyncpg). Ruff is also not installed in the local Python environment.